Merkelbach Paper

Last updated Feb 27, 2024 Edit Source

• Paper notes

#work/patientsim #paper_notes

Based on the MIMIC dataset.

• time series data collected, normalized
• filled in missing data, imputed via linear interpolation, or in the case of no intag:#work/patientsimstance of this data at all, the median
• static data tacked onto every single instance of the time series data

(it’s bidirectional) processes the whole of the data in both directions. Outputs are averaged over all the time steps, presumable leaving a single vector the length of the number of features.

The single vector is put through a dense layer, making the bottleneck. That bottleneck is “repeated through time”, which I just think means it’s copy/pasted as many times as there are timesteps, with the related timestep features added on.

Takes in the expanded bottleneck, turns it into another set (size of n timesteps) of vectors. A dense layer turns that into a copy of the original data.

Pasted image 20240125124146.png

• data is imputed - I’d personally rather it be robust to missingness
• static data repeated too much. We shouldn’t have to repeat ourselves.
• the gru is just summed. what a weird way to combine all those values. Really messy. There’s gotta be a better way.
• really, why is a gru even used?? Simply because it’s the only thing that could fit a time series?

Triplet Sparsity Reduction is my proposed solution to those issues.